Summary

Large models enabling algorithms to self-evolve is becoming a new paradigm. Two weeks ago this month, Tao Zhexuan and several mathematicians used Google’s AlphaEvolve, released in May this year, to solve Erdős 1026 with AI assistance after more than five decades of dormancy. Baidu’s 伐谋 follows a broadly similar idea; the split is that Google is betting on frontier mathematical research while Baidu is targeting industrial deployment—“this probably has something to do with the different national conditions in China and the US.”
Evolution is fundamentally a population-level process, not a single-Agent process—and that is where 伐谋 differs from the “self-iterating” Agents on the market. Having one Agent try 10 times and improve on the 11th “can at most be called memory or experience.” 诗婕 compares 伐谋’s population search with VC logic, while 安南 sees it as especially well suited to high-value algorithm problems: “We don’t care whether each Agent lives or dies. Let 100 Agents try simultaneously; if one produces a 100x return, the system works.” Large models handle “reproduction” between algorithms, while humans write the evaluator as the baton; in practice, “even if 80% of 1,000 evolutionary rounds are still reporting bugs, as long as 10% keep improving, the system works.”
Algorithms sit in high-value, mission-critical settings, which is why Robin presented 伐谋 himself. Unlike chat, customer service, or K12 problem solving, algorithms address core B2B decision and forecasting problems, where “a 1% or 5% improvement can mean a leap in output for the company”—optimizing quay-crane scheduling at a port could generate tens of millions in value. Robin put it plainly: “I’m presenting 伐谋 myself today because I think it matters.”
For algorithmic and research work, Gemini is more useful than Claude, with the turning point being the March experimental release of Gemini 2.5 Pro. 李安南’s explanation is that algorithm engineers need raw intelligence, while software engineers rely more on the experience of a veteran operator; Cursor-style long context is an engineering capability that algorithms do not necessarily require. Gemini can handle external validation work, while customer projects use Chinese models for security and privacy. The stack also mixes DeepSeek R1, Qianwen, Baidu’s own models and others—“the exact same Prompt and the exact same code produce different new outputs,” because every model has its own “algorithmic profile.”
The commercial model is still completely unsettled; the current KPIs are scale and ecosystem, not revenue. The team is testing three paths: RaaS charged on results, subscriptions built around the inherently ongoing nature of algorithm optimization, and a Palantir-style FDE deployed on-site. Attribution is a real problem: output gains at ports and factories are easy to calculate, while the returns from financial risk control and AI for Science are “indeed unknown.” ByteDance, Huawei and Alibaba’s cloud businesses do not yet have especially clear commercial products in this market; YC-backed startups have started to emerge overseas, while the domestic market has yet to produce an especially clear leader.
The structure is relatively resistant to hallucinations and can enter serious production: humans converge the process at both ends, while AI handles the middle. People define the problem, write the evaluator and perform the final review—the “human review and AI review work exactly the same way.” A local evaluation setup, with code generated in the cloud and tested locally, keeps data within the customer’s domain and reduces the need for large inventories of Kunlunxin, Nvidia or Ascend chips. The risks are just as real: in a port case, AI interpreted a rail-mounted system as a vehicle capable of crossing the gantry, exploiting an evaluation constraint that had not been fully specified. “The metric looked exceptionally good, but it was actually doing some very strange things.”
The change in production relations matters more than the productivity gain. Engineers can move from writing code to “one person guiding 10 AIs”; one automotive customer said it had 100 scenarios it wanted to transform, but previously lacked the people to do it. Deployment requires breaking the muscle memory of “locking orders for 3 days”—the wall “isn’t load-bearing, but it has always been there.” Companies will need internal “efficiency pioneers” to identify which walls can come down. 诗婕 extrapolates that many enterprises may eventually form cabinet-style structures with both conservatives and radicals; 安南 called the idea “brilliant.”

Deep dive

1. Tao Zhexuan solved Erdős 1026; what China missed was the AlphaEvolve in his hands

诗婕’s opening: two weeks ago this month, Tao Zhexuan and several mathematicians solved Erdős 1026 with AI assistance after the problem had remained untouched for more than five decades. The tool was AlphaEvolve, launched by Google in May this year—“an evolutionary algorithm-coding Agent based on large language models.” In China, little attention was paid to the tool itself.
安南 was born in 1998 and studied automation at Tsinghua from 2016 to 2023. His internships were in computer vision, during the era of SenseTime and Megvii—the “AI Four Little Dragons”—and autonomous driving. He entered the 2022 fall recruiting cycle just as ChatGPT launched, joined Baidu after graduation to work on B2B large-model applications, and is now 伐谋’s product lead.

2. A primer: algorithms have quality differences, unlike ordinary code

安南’s definition: an algorithm is mathematics at its core, with code as its implementation—“input a pile of information, pass it through a black box, and output more information.” The lineage runs from Turing-machine-era adders to sorting algorithms in the 1960s and 1970s, then to probability-based search algorithms in the 1970s and 1980s, where four directions are explored and the process iteratively moves toward a better one.
The key distinction is that “a worse algorithm may need to calculate 100 times, while a good algorithm may get the answer in just five.” It is not enough for the function to work and contain no bugs. Since the 1990s, the field has become data-driven, moving from SVMs to neural networks, deep learning, large models and Agents; machine learning “doesn’t care about the mechanism in the middle and learns features purely from data.”

3. The core proposition: use sufficiently advanced intelligence to optimize lower-level intelligence

The task is straightforward to state: “The input is my requirement—help me optimize a facial-landmark-recognition algorithm—and the output is the algorithm itself.” An Agent is used to optimize earlier-generation algorithms, from adders and sorting algorithms to machine learning.
The premise reflects a relatively mainstream industry view: “For problems that can be evaluated clearly, AI will probably solve them.” Evaluating facial landmarks is simple—where the eyes, nose and mouth are located is objectively observable—but writing the algorithm is difficult. Easy evaluation and difficult generation are precisely the territory targeted by AlphaEvolve and 伐谋.

4. AlphaEvolve did not appear out of nowhere: Fan Search, EoH and Baidu’s Prompt tuning last year

The lineage runs through DeepMind’s Fan Search, an earlier system that optimized specific algorithm fragments when large models were still weak, and EoH from City University of Hong Kong. Baidu also explored an Agent that “self-iterated based on an explicit evaluator” in December last year. Because coding ability was still limited, its main result was automatically improving the Prompt.
The turning point came this year. As large models’ coding capabilities improved materially, “having them automatically modify an algorithm became something that was very much called for.”

5. From AGI to ASI: the constraint is not human imagination but human time

Google has used AlphaEvolve for internal business applications, as disclosed in its paper. More importantly, it worked with mathematicians to discover new theorems. “Everyone is focused on Tao Zhexuan, without paying much attention to the fact that what Tao Zhexuan was using was Google AlphaEvolve.” 诗婕’s extension is that discovering results unknown to humanity could amount to reaching ASI; 安南 called the idea “very promising.”
The logic offered by Tao Zhexuan and his collaborators is that many problems are “not necessarily beyond what humans can think of; I simply don’t have that much time.” The tool “can amplify Tao Zhexuan’s capabilities by many multiples.” Baidu’s divergence is that some industrial problems could be solved by an engineer “if you gave him 100 years, but it’s impossible to make him calculate for 100 years.” They are suitable for handing to AI as compute scales up—“which probably has something to do with the different national conditions in China and the US.”

6. In algorithmic and research settings, Gemini beats Claude

安南 limits the comparison to a specific setting: “Whenever people talk about coding, they mention Claude—but in our research or algorithm scenarios, Gemini is actually the better model.” The difference comes down to the users: algorithm engineers need “raw intelligence,” tend to be younger and highly educated, and compete on ideas and insight; software engineers “feed on experience” and are veteran operators who have seen complex systems. Cursor’s prized long context is an engineering strength that algorithms do not necessarily need.
The turning point was the experimental version of Gemini 2.5 Pro in March. “We tested it at the time and it really was a significant improvement.” The explanation remains a hypothesis, not a conclusion: both models are closed-source, but DeepMind may focus more on research, with data and post-training tilted in that direction. Anecdotally, more university students use Gemini for algorithms and research because “it really is smarter.”

7. Customer work mixes Chinese models; every model has its own “algorithmic profile”

External validation and open-source work can be built on Gemini. Customer work uses Chinese models because of security and privacy requirements. The overall stack mixes DeepSeek R1, Baidu’s own models, Qianwen and others because different models produce different ideas and head in different directions.
The reason for mixing them is simple: “Give them the exact same Prompt and the exact same code, and what they generate will still be different.” 诗婕 summarizes that gap by saying each model has its own “algorithmic profile.”

8. Why Robin presented 伐谋 himself: algorithms sit in high-value scenarios

At the Baidu World Conference, Robin generally serves as host; 伐谋 was the exception. He launched it himself: “I’m presenting 伐谋 myself today because I think it matters, and I’m also passionate about expressing it.”
安南’s explanation is that previous B2B deployments largely involved chat, role-playing, K12 problem solving and e-commerce quality inspection—“important, but the value may not show up that quickly.” Algorithms address critical decision and forecasting problems across industries, where “a 1% or 5% improvement could mean a leap in output for the company.” Algorithm engineers are expensive precisely because they can create commensurate value.
The consumer-facing example is Didi’s vehicle dispatch, itself a large algorithm. “Without algorithmic control, we would go back to calling drivers by phone in the 1990s, and the business model wouldn’t exist.” Replace Didi with a company in any other industry and ride-sharing with any other sector, and the logic holds: “Wherever algorithms sit is where the high-value scenarios are.”

9. Evolution is a population concept: from startup trial and error to VC logic

安南’s distinction from “self-evolving Agents” is direct: “I tried 10 times, gained some experience, and did better on the 11th—that can at most be called memory, at most experience. It still cannot be called self-evolution.” Evolution must be a population-level concept.
诗婕 compares 伐谋 with VC logic: it does not care whether each Agent succeeds or fails; 100 Agents try different approaches simultaneously. As long as the portfolio odds are right, one Agent producing a 100x return makes the overall system work. 安南 believes this mechanism is particularly suited to algorithm problems: there is no need to optimize the compute used by each individual Agent, so the system can run 16-way concurrency and conduct a much larger search.

10. Evolutionary algorithms × large models: islands, reproduction, evaluators and oil drilling

An evolutionary algorithm is Darwinian: 5 islands hold 10 individuals each; better performers receive more opportunities to reproduce; after 10, 100 or 1,000 generations, selection and elimination take place, with crossover between islands, gradually moving toward a global optimum. In the past, mechanically combining two algorithms would trigger bugs. Now “the process of having children is handled by the large model”—it uses A and B to generate a better C. Humans decide what survives by writing the “evaluator,” which functions “a bit like a baton.”
The evolutionary strategy can be understood as drilling for oil: “When should you drill this well a little deeper, and when should you give up and drill a new well somewhere else?” Staring at one well and continuing to dig risks getting trapped in a local optimum. That is the difference between an individual and a population.
诗婕’s analogy was well received: microorganisms exposed to different environments evolve into apes, deep-sea fish or dinosaurs; some survive and some are eliminated. 安南 responded: “It doesn’t matter who survives. As long as one of them ends up being good at something, we take that thing out.”

11. It still works if 80% of the runs report bugs: probability is the decisive edge

Large models contribute probability. If coding ability is strong, the probability of successfully producing the next generation exceeds 50%; “multiply that by 1,000 trials and the thing works.” If it is below 50%, evolution deteriorates and even 10,000 rounds will not produce a result. The practical observation is: “Maybe 80% of 1,000 evolutionary rounds are still reporting bugs, but as long as 10% are continuously improving, 伐谋 works.”
The work of algorithm engineers across industries can be abstracted into 4 steps: select a baseline, read papers to find a new idea, generate a new algorithm based on the idea and the foundational algorithm, then run a validation set and review cases before repeating the cycle. All 4 steps can be turned into Agent workflows. For humans, the more important tasks are defining the problem and setting the evaluator—the “baton”—correctly.

12. “Shangbing Famu”: algorithmic generality versus the limits of vertical fine-tuning

The name comes from The Art of War’s “Shangbing Famu”: “mou” means strategy, while “fa” means to attack or overcome. An algorithm is also, fundamentally, a strategy.
A counterexample came from a presentation at Tsinghua’s Department of Electrical Engineering. Students vertically trained a power-dispatch model and improved its performance, but its “generalizability was extremely poor”; moving to logistics or port scheduling meant starting over. 伐谋 writes the algorithm itself, making it more transferable across industries—“the same more general Agent can serve scheduling scenarios in industries A, B, C and D, rather than doing something highly vertical for each industry.”
This contrasts with last year’s Huawei Cloud-style path of deploying fine-tuned models in vertical scenarios. Vertical large models have value, “but it is difficult to truly serve every industry; the process is very slow, and there are also data problems.” 伐谋 wants to build an algorithm-Agent ecosystem in a more general form.

13. Forecasting and decision algorithms run through the same 4-step loop

In forecasting, a retail customer wants to know roughly how many units will sell tomorrow, the day after tomorrow and the day after that, using a large volume of historical and competitor data that might previously have been handled by gut feel. In decision-making, once demand for 100 orders is known, the system must decide when to add vehicles and which vehicles to dispatch; a large SKU base and limited fleet make the problem highly complex. Didi is primarily a decision algorithm.
The underlying framework is the same for both: they return to the 4-step loop. “That is why they can be upgraded with AI and become more general,” bringing capabilities previously available only to algorithm teams within reach of mid-sized companies.

14. B2B deployment in the field: “AI used to do the work of liberal-arts students; now it is doing the work of STEM students”

Customers’ first reaction is skepticism: “AI can really do something this difficult?” But the cost of trial is low. Algorithms are abstracted, self-contained modules that do not deeply penetrate business processes; a basic scenario can be built in a month or less, enough to make customers happy or surprised. One large central state-owned enterprise said: “It feels like most of what AI did before was the work of liberal-arts students, but now we suddenly realize that 伐谋 is doing the work of STEM students.”
The structure is relatively resistant to hallucinations because the objective is clear and the evaluator is objective: “The way a human reviews and the way AI reviews are exactly the same.” Humans define the objective and conduct the final evaluation, leaving the middle to the large model.
The bigger change is continuity. Previously, algorithm engineers worked on projects: after a 5% improvement went live, the project ended, and it was impossible to bring the experts back when the problem resurfaced. With AI, the process only needs to be restarted. “It is no longer just a productivity-enhancement logic; it is more like a change in the productive forces or the production relations.” Policy has also called for “autonomous, continuous optimization” across industries.

15. Agent infra: evolutionary operators plus distributed scheduling compress a week into 2 or 3 days

A single Agent does not need much complex infrastructure; “the large model’s infra is already sufficient for it.” 伐谋’s infra has 2 parts: the core operators that steer the evolution—the “construction-site foreman” when drilling for oil—and the underlying infrastructure. With serial APIs waiting on one another, 1,000 iterations take 5 days to a week. Distributed parallelism, step decomposition and fuller use of idle GPU and CPU capacity can bring that back to 2 or 3 days.
The organizational model is “a small team inside a large company.” A large company is necessary because the system needs an ecosystem—university labs, startups and vertical champions working together to serve every industry—as well as long-term accumulation in cloud resources and scheduling capabilities. The team must stay small because the model is still early: it cannot be rolled out wholesale according to a fixed template and instead requires continuous trial, validation and replication.
Startups face a harder time winning customer trust. The product is too new and customers are already skeptical, while Baidu’s AI brand gives them “a sense of security and a backstop.”

16. The commercial model is still wide open: RaaS, subscriptions and Palantir-style FDE

There are 3 paths. RaaS, or “Result as a Service,” charges for results—for example, if algorithm optimization saves RMB5M, the fee is based on that outcome. “It is relatively new in China today, but may be a trend in the AI era.” The second is a subscription model: Chinese SaaS has never truly taken off, but “algorithm optimization itself is an ongoing process,” akin to hiring 10 engineers who must create value every year. The third is a Palantir-style forward-deployed engineer stationed at the customer site: “a person takes 伐谋 and solves the problem there.”
诗婕 identified the attribution problem: it is difficult to determine how much of a customer’s improvement came from the algorithm. 安南 acknowledged it: “At ports and in manufacturing, it is easy to calculate the output and revenue, but for AI for Science or financial risk control, what exactly is my return? That is indeed unknown.” The model is therefore still being explored.

17. 伐谋 also has a consumer side: university students who probably will not pay are a 5-to-10-year variable

Responding to 李开复’s 2023 view that China’s AI would move to consumers while the US would go to B2B first, 安南 sees it as a question of timing. Consumer value is larger over the long term but difficult to validate commercially in the short term; in B2B, choosing the right high-value scenario produces a clear willingness to pay.
An unexpected discovery was that, after the platform launched, many university students began using it for coursework, graduation projects and papers. A student with a fluid-mechanics background could use it to build a simulation model—an AI problem—without actually learning AI. “We probably will not charge this group; it is more like cultivation.” Once these students enter the workforce, they may naturally assume that algorithm problems should be solved with AI.
The team itself reflects that profile. The 1998-born 安南 leads campus hires and interns born around 2000, many of whom come from operations research, industrial engineering and the sciences rather than from AI.

18. Customer sentiment and local evaluation: keeping data within the domain is the China solution

The contrast with 2023 is stark. Back then, the imperative was “get it deployed quickly, cheaply and efficiently—buy chips, buy DeepSeek models and deploy a whole lot of them.” Now the pain point is “how do we actually use all this?” Customers are especially pleased when already-deployed models can be used in “very niche, very valuable scenarios.” The 2025 inference era and improved coding ability restored confidence; after the release of cases in financial risk control, port scheduling and China Energy Engineering’s route planning, “customers with very similar needs all started coming to us.”
The technical answer is local evaluation. The compute-intensive code-generation step runs in the cloud, while evaluation runs on the customer’s local data and only the final result is sent back. “The data does not need to leave the domain, and we do not need to prepare so many Kunlunxin, Nvidia or Ascend chips.” It reflects the different data and deployment conditions of industrial and research settings.
The old problem of unrealistic feedback from central and other state-owned enterprises is real and cannot be solved technically. But the product is new enough, and policy support strong enough, that trials can escape some of the constraints imposed by customer relationships and other factors. Mid-sized privately owned vertical champions care more about results and are more likely to provide honest feedback: “They are more anxious than we are; right after the conversation, they rush to send over the sample data for testing.”

19. Intelligent-native operations: breaking walls, locked orders and Delta algorithms

The first layer of “internalizing AI” is breaking human priors. Orders are classified from P1 to P4 and processed sequentially because people cannot handle global optimization. AI can break up that sequence and search globally across gross margin, revenue, delivery times and elasticity. 安南’s analogy is: “There is a wall between the bedroom and the living room. It is not load-bearing and could actually be knocked down, but it has always been there.” Removing it can require communication and even growing pains; in manufacturing, the rule of “locking orders for 3 days” could be shortened to 1 day once AI is fast enough.
The second layer is continuously searching for Delta algorithms. Particularly busy and particularly idle operating conditions may require algorithm A and algorithm B. “The more general an algorithm is, the less impressive its vertical performance may be.” 伐谋 continuously searches for the best solution for each operating condition as new data flows in. The product’s 3 defining traits follow: generality, industrial-grade capability and continuity.

20. Product form: not a chatbot, but a creator watching the evolutionary family tree

The product has 3 stages. First, users submit a task conversationally; it began as a web product and will also adapt to different workflows such as VS Code and Notebook, as long as the task and evaluator are clearly specified. Second, users monitor an evolutionary family tree: “It is a bit like we are the creator, watching how far each organism below has evolved.” Tree nodes that report bugs or score poorly turn gray and are eliminated; active nodes brighten and receive more chances to iterate. Third, the audit stage explains how the best algorithm evolved step by step, providing interpretability.
It cannot simply be a chatbot because “if I lead 10 interns to complete an algorithm task, it is hard to have conversations with all 10 every day.” Humans are mainly responsible for monitoring, review and the final explanation, not for holding one conversation with every Agent. 诗婕’s closing image is that the user plants one seed, the large model turns it into hundreds or thousands of seeds, they are cultivated on different terrain, and the system monitors them and selects the winners.

21. God and trouble: surprises in factor mining, and a port’s rail-system hack

The positive side is visible in financial risk control. The feature extractors and factor mining produced by AI “were genuinely things our human algorithm engineers had not really thought of before.” On inspection, they made sense and provided useful inspiration.
The troubling side is “still quite like a science-fiction movie in the long run.” Handing traffic lights, power grids or port scheduling to AI carries risk and requires strict auditing. In a real case, AI found a scheduling plan that used very few machines to complete a large number of tasks. On closer inspection, the evaluator’s constraints were incomplete: it treated a rail-mounted piece of equipment as a vehicle capable of crossing the gantry, even though the real equipment has physical rail relationships. “The metric looked exceptionally good, but it was actually doing some very strange things.”
The conclusion returns to humans. “The amount of intelligence is proportional to the amount of human input” remains true in the large-model era—and that is a good thing. “If it weren’t, that would actually be the real doomsday scenario. This absolutely cannot be about replacing people; in all likelihood, someone will still need to define the problem, audit it and monitor it.”

22. Production relations matter more than productive forces: one person directing 10 AIs, and the corporate “cabinet”

The change in production relations is that algorithm engineers can hand coding, testing and exploration to AI and become “one person guiding 10 AIs.” This is not simply a matter of cutting 9 people: one automotive company said, “I have 100 scenarios I actually want to transform; previously I just didn’t have enough people.” The new profile is someone who understands the business and knows a little about algorithms, rather than a two-person pairing of a business expert and an AI expert.
The new “efficiency pioneer” is tasked with finding which walls inside the company can be knocked down. The role resembles a Palantir FDE, but does not necessarily require an AI background; it is more like an entrepreneur who knows how to use AI tools. The role is “very much worth looking forward to” and could help drive change from within the enterprise.
诗婕 therefore expects many companies to establish a “cabinet system” with both conservative and radical factions. 安南 responded: “Brilliant.” In 5 to 10 years, consumers may simply feel that efficiency is higher, while B2B R&D teams may look more like business partners, defining problems with operating teams and driving AI proactively.

23. Closing: the KPIs are scale and ecosystem; what is being launched are cases, not stories

The priority today is “not commercial revenue—it is too early,” but the scale and ecosystem of 伐谋. 伐谋 is only one key Agent on 千帆’s enterprise-Agent infra and AI-infra platform. Its evolutionary capabilities, logic, performance and scheduling can be reused by other self-evolving Agents. Under integrated cloud and intelligence, the compute consumed when a customer runs 100 concurrent Agents—including domestic chips such as Kunlunxin—can be bundled with cloud services.
The competitive scan: ByteDance currently looks more like a public-cloud play and has relatively few offerings this vertical. Huawei is moving quickly, but “the core logic still relies on human algorithm engineers,” leaving it oriented toward projects. Alibaba is also exploring the technology and related supporting systems. None of the major cloud providers yet has an especially clear commercial product in this market; a wave of YC-incubated startups has emerged overseas, while China has not yet produced an especially clear one.
The 4 priority sectors are energy, finance, manufacturing and consumer, which 安南 estimates together account for “more than 30% of China’s GDP, perhaps even more.”
伐谋 began as an internal project to improve efficiency for Baidu Cloud, finance, transportation and other divisions. At the World Conference, “rather than launching a product, we were essentially launching several cases.” The team “cannot keep talking about future stories; the fruit has to grow as quickly as possible.” Robin’s line—that projects where technology plays a decisive role are acceptable even if 9 out of 10 attempts fail—resonates internally. Baidu chose the harder, technology-led path; in B2B, “the wine smells good,” but the process has to go deeper.

Where Did the Money Go? Who’s Making Moves? A Review with 刘旌 of VCs’ New Consensus and Undercurrents in 2025 Controlling Tools or Aligning Creatures? Emmett Shear (Softmax) & Séb Krier (GDM), from a16z Show